I'm training a sentiment classifier based on DNN. Currently my training set is made out of 45% positive sentiments, 45% negative sentiments and only 10% of neutral sentiments. As you probably guess, the precision I get is low, especially for the neutrals.
I'm looking for a source where I can create such neutral sentiment corpus. I was thinking about articles but it's too literature language while I'm looking for more 'social language' sentences. I came to think about using Wikipedia's corpus as my neutrals corpus but still I'm not sure it belongs to the profile of 'social language' that I'm looking for. Facebook & Twitter are out of question because the sentences may be positive, negative or neutral.
I'm looking for suggestions for my problem.